Skip to content

Directives Reference

This page lists every directive in AIA 2.0 in alphabetical order, with its aliases, arguments and an example. For an introduction, read Directives first.

Syntax

Where Form Notes
Chat /name arg1 arg2 Must be on a line by itself. The name is case-insensitive. Arguments are split on spaces, and quotes don't group words.
Prompt file <%= name 'arg1', 'arg2' %> An ERB tag. Its result replaces the tag. A directive with no arguments is just <%= name %>.

Use in shows where each directive is meant to be used:

  • Chat: designed for interactive sessions.
  • Prompt: returns text, so it is useful inside a prompt file.
  • Both: useful in either place.

Chat output describes what happens to the result in a chat session:

  • Sent: AIA prints the result and also sends it to the model as part of the conversation.
  • Shown: AIA prints the result only. Nothing reaches the model.
  • Next prompt: AIA turns on a mode that applies to your next message only.

Summary

Directive Aliases Use in Chat output Purpose
/add_recruit /add Chat Shown Add a persistent robot to the crew
/available_models /am, /available, /models, /all_models, /llms Chat Shown Search the model list
/checkpoint /ckp, /cp Chat Shown Save a point in the conversation
/checkpoints_list /checkpoints Chat Shown List checkpoints
/clear Chat Shown Forget the conversation
/compare /cmp Chat Shown Ask several models the same thing
/concurrent /conc Chat Next prompt Query MCP server groups in parallel
/config /cfg Chat Shown Show or change settings
/cost Chat Shown Token and cost report as CSV
/debate Chat Next prompt Multi-round debate between robots
/decompose Chat Next prompt Split a prompt into parallel sub-tasks
/delegate /del Chat Next prompt Assign sub-tasks to robots via TrakFlow
/drop_recruit /drop Chat Shown Remove a robot from the crew
/help Chat Shown List all directives
/include Both Sent Insert another file, processed as a prompt
/insert /read Both Sent Insert a file's raw text
/mcp Chat Shown MCP server status and tools
/model Chat Shown Show or switch the model
/orchestrate /orch Chat Next prompt Three-tier orchestration
/paste /clipboard Both Sent Insert the clipboard contents
/plan Chat Sent Create a TrakFlow plan
/plugins Chat Shown List loaded plugins
/reskill Chat Shown Reset a crew member with new skills
/restore Chat Shown Go back to a checkpoint
/review /context Chat Shown Show the conversation
/robot Chat Shown ASCII robot art
/robots Chat Shown Show the active robot(s)
/ruby /rb Chat Sent Run Ruby (needs --allow-ruby-eval)
/say Chat Shown Speak text aloud
/skill Both Sent Insert a skill's text
/skills Chat Shown List skills
/spawn Chat Next prompt One-off specialist robot
/task Chat Sent Create a TrakFlow task
/tasks /tf Chat Sent Show TrakFlow ready tasks
/temperature /temp Chat Shown Set temperature
/tools Chat Shown List available tools
/top_p /topp Chat Shown Set top-p
/verify Chat Next prompt Two answers plus reconciliation
/webpage /website, /web Both Sent Insert a web page as Markdown

Front matter keys such as next: and pipeline: aren't directives. See Prompt Files. There are no /next or /pipeline chat directives.

/add_recruit

Aliases: /add · Use in: Chat · Chat output: Shown

/add_recruit NAME [PROVIDER/MODEL] [skill:ID[,ID...]] [SYSTEM PROMPT...]

Adds a robot named NAME to the crew. A recruit stays for the rest of the session, and you can address it with @NAME.

  • With only a name, the recruit uses the session's model.
  • Use - or inherit in place of a model to keep the session's model while still giving a system prompt.
  • Each skill:ID token assigns one or more skills. The skills become the recruit's role.
  • Everything after the model is the recruit's system prompt.
/add_recruit critic openai/gpt-4o You are a blunt code reviewer.
/add tester - skill:rspec_writer

AIA replies Recruited 'critic' into the crew (openai/gpt-4o). If it can't add the recruit, it replies Recruit failed: .... See Crews.

/available_models

Aliases: /am, /available, /models, /all_models, /llms · Use in: Chat · Chat output: Shown

/available_models [TERM ...]

Lists models with their provider, input price per million tokens, context window, modalities and capabilities.

  • A plain term (or +term) keeps the models whose entry contains it.
  • -term, ~term or !term removes the models whose entry contains it.
  • A term with _to_, such as text_to_image, filters by modality.
  • If your current model has a provider/ prefix, AIA asks that provider for its live model list instead.
/models claude -haiku
/am text_to_image

/checkpoint

Aliases: /ckp, /cp · Use in: Chat · Chat output: Shown

/checkpoint [NAME]

Saves the current conversation under NAME. Without a name, checkpoints are numbered 1, 2, 3 and so on. Checkpoints last only for the current session.

/checkpoint before_refactor
Checkpoint 'before_refactor' created at position 6.

/checkpoints_list

Aliases: /checkpoints · Use in: Chat · Chat output: Shown

Lists the saved checkpoints with their position, the time each was created, and a preview of the last message you sent before it.

/checkpoints

=== Available Checkpoints ===
  before_refactor: position 6, created 14:02:11
    → "Now refactor the parser to use a state machine"
=== End of Checkpoints ===

/clear

Use in: Chat · Chat output: Shown

/clear [--all]

Removes the conversation history from every robot in the session. The system prompt is kept unless you add --all. All checkpoints are deleted too.

/clear
Chat context cleared.

/compare

Aliases: /cmp · Use in: Chat · Chat output: Shown

/compare PROMPT --models MODEL1,MODEL2[,...]

Sends the prompt to each listed model separately and prints each response under the model's name. The models don't see your conversation.

Every word that isn't part of --models belongs to the prompt.

/compare Write a haiku about autumn --models gpt-4o-mini,claude-haiku-4-5

/concurrent

Aliases: /conc · Use in: Chat · Chat output: Next prompt

Runs your next message with your MCP servers split into independent groups that are queried in parallel. This needs more than one MCP server that is relevant to the prompt, in at least two groups. Otherwise the message runs normally. To do this automatically on every prompt, use --concurrent-auto. See MCP Servers.

/concurrent
Compare the open GitHub issues with the TODOs in the local repository.

/config

Aliases: /cfg · Use in: Chat · Chat output: Shown

/config                  # show every setting
/config SECTION          # show one top-level section, e.g. llm, flags, output
/config KEY VALUE        # change a setting

KEY can be:

  • An option name, such as temperature, max_tokens, debug, verbose, consensus or tokens
  • A dotted path, such as llm.max_tokens or output.file

The change lasts for the rest of the session only. Model settings such as temperature, top_p and llm.* apply from your next message: AIA retunes the running robot, and every crewmate, without losing the conversation. See Configuration Precedence. AIA warns Unknown config option 'x' if it doesn't recognize the key.

/config llm
/config debug true
/config llm.max_tokens 4000

Value types

AIA converts the value to the setting's type: numbers for temperature or llm.max_tokens, true/false for on/off settings (true, yes, on, 1 / false, no, off, 0), and comma-separated lists for list settings. A value that doesn't fit is stored as you typed it.

/cost

Use in: Chat · Chat output: Shown

Prints a CSV table of the session's turns. It shows the model, input, output and total tokens, the cost and the elapsed time for each turn, followed by a TOTAL row. If an output file is set, the CSV is also appended to it.

/cost

model,input_tokens,output_tokens,total_tokens,cost,elapsed
gpt-4o-mini,412,256,668,$0.00022,2.1s
TOTAL,412,256,668,$0.00022,2.1s

/debate

Use in: Chat · Chat output: Next prompt

Your next message becomes the subject of a debate among the robots in the session. The debate runs at least 2 rounds and at most 5, and stops early when the robots agree. You need at least two robots, from several models (--model a,b) or from recruits. With only one robot, the message runs normally. See Advanced Modes.

/debate
Should a small team start with a monolith or microservices?

/decompose

Use in: Chat · Chat output: Next prompt

Splits your next message into independent sub-tasks, runs them in parallel, and combines the results. Timing figures show how long the work took in parallel compared with one after another. If the message can't usefully be split, it runs normally.

/decompose
Summarize the pros and cons of Postgres, MySQL and SQLite for a mobile backend.

/delegate

Aliases: /del · Use in: Chat · Chat output: Next prompt

Breaks your next message into sub-tasks, assigns each one to a robot in the crew, records the plan in TrakFlow, and runs it. This needs more than one robot and an initialized TrakFlow project (tf init). Otherwise the message runs normally.

/delegate
Design the database schema, write the migration, and write tests for it.

/drop_recruit

Aliases: /drop · Use in: Chat · Chat output: Shown

/drop_recruit NAME

Removes a recruit from the crew.

/drop critic
Dropped 'critic' from the crew.

/help

Use in: Chat · Chat output: Shown

Lists every directive, including custom ones, grouped by category with their aliases and descriptions.

/include

Use in: Both · Chat output: Sent

/include PATH                       # chat
<%= include 'PATH' %>               # prompt file

Inserts another file after processing it as a prompt. Its shell references, ERB and parameters are all handled, and it can include other files in turn. In a prompt file, PATH is relative to the directory of the file that contains the tag. In chat, it is relative to your current directory. Include loops are detected and stop with an error. include isn't available in executable or piped prompts. Use insert there.

<%= include 'shared/house_style.md' %>

/insert

Aliases: /read · Use in: Both · Chat output: Sent

/insert PATH
<%= insert 'PATH' %>

Inserts a file's text exactly as it is, with no shell expansion, ERB or parameters. Paths resolve the same way as for include, or against the current directory when there is no prompt file.

Review this code:

<%= insert 'lib/parser.rb' %>

/mcp

Use in: Chat · Chat output: Shown

Shows how many MCP servers are defined, connected and failed. It lists each connected server's tools and the error for each failed server.

/mcp

MCP Server Status
=================
Defined: 2  Connected: 1  Failed: 1

Connected Servers:
  github (26 tools)
    - list_issues
    ...
Failed Servers:
  postgres: connection refused

/model

Use in: Chat · Chat output: Shown

/model                       # show details of the current model(s)
/model MODEL[,MODEL...]      # switch to one or more models

With no argument, AIA shows the current model's details. With several models, it shows each one's provider, context window, prices and capabilities.

Switching keeps your conversation: by default the old model summarizes it for the new one. Set model_switch_history to replay (re-send your messages) or clean (start fresh). MODEL can also be an alias such as sonnet or cheap, or one you define under model_aliases.

See Choosing Models.

/model claude-sonnet-4
/model gpt-4o,claude-sonnet-4

/orchestrate

Aliases: /orch · Use in: Chat · Chat output: Next prompt

Treats your next message as a set of application requirements. Three tiers of robots work on it: an orchestrator, lead agents and specialists. See Advanced Modes.

/orchestrate
Build a command-line tool that tracks reading habits, with a SQLite store and a weekly report.

/paste

Aliases: /clipboard · Use in: Both · Chat output: Sent

Inserts the text on your system clipboard.

/paste

Fix the grammar in this paragraph:

<%= paste %>

/plan

Use in: Chat · Chat output: Sent

/plan DESCRIPTION

Creates a TrakFlow item from the description. This needs an initialized TrakFlow project (tf init). Otherwise AIA prints TrakFlow not available. Run 'tf init' to initialize a project.

/plan Migrate the reporting jobs from cron to Solid Queue

/plugins

Use in: Chat · Chat output: Shown

Lists the plugins that are currently loaded, by file name without .rb.

/reskill

Use in: Chat · Chat output: Shown

/reskill NAME [skill:ID[,ID...]] [SYSTEM PROMPT...]

Clears a crew member's conversation and gives it new skills, a new system prompt, or both.

/reskill critic skill:security_review
Reskilled 'critic' with security_review (clean slate).

/restore

Use in: Chat · Chat output: Shown

/restore [NAME]

Returns the conversation to a checkpoint. Everything after it is discarded, and so are any checkpoints saved after it. Without a name, AIA restores the checkpoint before the most recent one. That takes at least two checkpoints.

/restore before_refactor
Context restored to checkpoint 'before_refactor' (position 6).

/review

Aliases: /context · Use in: Chat · Chat output: Shown

Prints the conversation so far, one numbered line per message, with 📍 markers where checkpoints were saved. Long messages are shortened, and tool calls are summarized.

/robot

Use in: Chat · Chat output: Shown

Prints AIA's ASCII robot.

/robots

Use in: Chat · Chat output: Shown

Shows the active robot or robots and the mode: Single, Crew, Parallel, Pipeline or Consensus. For each robot it shows the model and provider, the cost per 1K input and output tokens, how many local and MCP tools it has, and its role.

/ruby

Aliases: /rb · Use in: Chat · Chat output: Sent

/ruby RUBY CODE

Evaluates one line of Ruby and sends the result to the model. Disabled by default. Enable it with --allow-ruby-eval, or with flags: { allow_ruby_eval: true } in ~/.config/aia/aia.yml. While it is disabled, the directive only prints an error. While it is enabled, AIA prints a warning to STDERR before it runs the code. If the code raises an error, the error message is sent in place of a result.

/ruby Dir.glob('spec/**/*_spec.rb').size

See ERB & Shell and Security.

/say

Use in: Chat · Chat output: Shown

/say TEXT

Speaks the text with your speak command (--speak-command or audio.speak_command; the macOS say command by default). With say, AIA uses your voice (--voice or audio.voice) if you've set one. Without text, AIA prints Usage: /say TEXT and sends nothing to the model. To have every response spoken, use --speak. See Images & Speech.

/say The build is green

/skill

Use in: Both · Chat output: Sent

/skill ID
<%= skill 'ID' %>

Inserts the SKILL.md text of a skill from your skills directory (default ~/.prompts/skills). ID can be the start of a skill's directory name, or a path that begins with /, ./, ../ or ~/, to a skill folder or to a .md / .markdown file. See Roles & Skills.

/skill changelog

/skills

Use in: Chat · Chat output: Shown

/skills [TERM ...]

Lists the skills in your skills directory with their name and description. Terms filter the list by the text of each skill's front matter: a plain term must appear, and a -term must not.

/skills ruby -rails

/spawn

Use in: Chat · Chat output: Next prompt

/spawn                                       # AIA picks a specialist for the next prompt
/spawn TYPE                                  # a specialist of this kind
/spawn NAME PROVIDER/MODEL [SYSTEM PROMPT]   # exactly the specialist you describe

Creates a one-off specialist robot to answer your next message. Unlike a recruit, a spawned specialist doesn't join the crew. In the full form, use - or inherit as the model to keep the session's model, and add skill:<id> to give the specialist a skill. A specialist gets the chief's tools and MCP servers. See Crews.

/spawn security
Review this login controller for vulnerabilities: <%= insert 'app/controllers/sessions_controller.rb' %>

/task

Use in: Chat · Chat output: Sent

/task TITLE

Creates a TrakFlow task with the given title. Like /plan, it needs an initialized TrakFlow project.

/task Add retry logic to the webhook sender

/tasks

Aliases: /tf · Use in: Chat · Chat output: Sent

/tasks             # ready tasks
/tasks summary     # project summary

Shows TrakFlow tasks that are ready to work on, or a summary of the project.

/temperature

Aliases: /temp · Use in: Chat · Chat output: Shown

/temperature VALUE

Sets the sampling temperature for the rest of the session. Lower values give more focused answers, and higher values give more varied ones. This is the same as /config temperature VALUE. In 2.0.0 the robot that is already running keeps its old temperature. The new value applies after you switch with /model, and to robots built later in the session.

/temp 0.2

/tools

Use in: Chat · Chat output: Shown

/tools [FILTER]

Lists the tools the model can call, both local and MCP, with their descriptions. FILTER keeps only the tools whose name or description contains that text. See Tools.

/tools file

/top_p

Aliases: /topp · Use in: Chat · Chat output: Shown

/top_p VALUE

Sets top-p (nucleus) sampling for the rest of the session. This is the same as /config top_p VALUE. Like /temperature, it doesn't change the robot that is already running until you switch with /model. Some providers reject requests that set both temperature and top-p.

/verify

Use in: Chat · Chat output: Next prompt

Gets two independent answers to your next message, then reconciles them into one answer. This is useful for questions where a single answer might be confidently wrong.

/verify
How many weekdays are there between 2026-11-02 and 2026-12-25 inclusive?

/webpage

Aliases: /website, /web · Use in: Both · Chat output: Sent

/webpage URL
<%= webpage 'URL' %>

Fetches a web page through the pure.md service and inserts it as Markdown. This requires the PUREMD_API_KEY environment variable. Without it, AIA prints ERROR: PUREMD_API_KEY is required in order to include a webpage.

/web https://docs.ruby-lang.org/en/master/NEWS_md.html

See also